2 * Note: this file was generated by the Gromacs avx_256_double kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_256_double.h"
34 #include "kernelutil_x86_avx_256_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_VF_avx_256_double
38 * Electrostatics interaction: ReactionField
39 * VdW interaction: CubicSplineTable
40 * Geometry: Water4-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_VF_avx_256_double
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
63 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
64 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
66 real *shiftvec,*fshift,*x,*f;
67 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
69 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 real * vdwioffsetptr0;
71 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 real * vdwioffsetptr1;
73 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
74 real * vdwioffsetptr2;
75 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
76 real * vdwioffsetptr3;
77 __m256d ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
78 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
79 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
80 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
81 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
82 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
83 __m256d dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
84 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
87 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
90 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
91 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
93 __m128i ifour = _mm_set1_epi32(4);
94 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
96 __m256d dummy_mask,cutoff_mask;
97 __m128 tmpmask0,tmpmask1;
98 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
99 __m256d one = _mm256_set1_pd(1.0);
100 __m256d two = _mm256_set1_pd(2.0);
106 jindex = nlist->jindex;
108 shiftidx = nlist->shift;
110 shiftvec = fr->shift_vec[0];
111 fshift = fr->fshift[0];
112 facel = _mm256_set1_pd(fr->epsfac);
113 charge = mdatoms->chargeA;
114 krf = _mm256_set1_pd(fr->ic->k_rf);
115 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
116 crf = _mm256_set1_pd(fr->ic->c_rf);
117 nvdwtype = fr->ntype;
119 vdwtype = mdatoms->typeA;
121 vftab = kernel_data->table_vdw->data;
122 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
124 /* Setup water-specific parameters */
125 inr = nlist->iinr[0];
126 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
127 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
128 iq3 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+3]));
129 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
131 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
132 rcutoff_scalar = fr->rcoulomb;
133 rcutoff = _mm256_set1_pd(rcutoff_scalar);
134 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
136 /* Avoid stupid compiler warnings */
137 jnrA = jnrB = jnrC = jnrD = 0;
146 for(iidx=0;iidx<4*DIM;iidx++)
151 /* Start outer loop over neighborlists */
152 for(iidx=0; iidx<nri; iidx++)
154 /* Load shift vector for this list */
155 i_shift_offset = DIM*shiftidx[iidx];
157 /* Load limits for loop over neighbors */
158 j_index_start = jindex[iidx];
159 j_index_end = jindex[iidx+1];
161 /* Get outer coordinate index */
163 i_coord_offset = DIM*inr;
165 /* Load i particle coords and add shift vector */
166 gmx_mm256_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
167 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
169 fix0 = _mm256_setzero_pd();
170 fiy0 = _mm256_setzero_pd();
171 fiz0 = _mm256_setzero_pd();
172 fix1 = _mm256_setzero_pd();
173 fiy1 = _mm256_setzero_pd();
174 fiz1 = _mm256_setzero_pd();
175 fix2 = _mm256_setzero_pd();
176 fiy2 = _mm256_setzero_pd();
177 fiz2 = _mm256_setzero_pd();
178 fix3 = _mm256_setzero_pd();
179 fiy3 = _mm256_setzero_pd();
180 fiz3 = _mm256_setzero_pd();
182 /* Reset potential sums */
183 velecsum = _mm256_setzero_pd();
184 vvdwsum = _mm256_setzero_pd();
186 /* Start inner kernel loop */
187 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
190 /* Get j neighbor index, and coordinate index */
195 j_coord_offsetA = DIM*jnrA;
196 j_coord_offsetB = DIM*jnrB;
197 j_coord_offsetC = DIM*jnrC;
198 j_coord_offsetD = DIM*jnrD;
200 /* load j atom coordinates */
201 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
202 x+j_coord_offsetC,x+j_coord_offsetD,
205 /* Calculate displacement vector */
206 dx00 = _mm256_sub_pd(ix0,jx0);
207 dy00 = _mm256_sub_pd(iy0,jy0);
208 dz00 = _mm256_sub_pd(iz0,jz0);
209 dx10 = _mm256_sub_pd(ix1,jx0);
210 dy10 = _mm256_sub_pd(iy1,jy0);
211 dz10 = _mm256_sub_pd(iz1,jz0);
212 dx20 = _mm256_sub_pd(ix2,jx0);
213 dy20 = _mm256_sub_pd(iy2,jy0);
214 dz20 = _mm256_sub_pd(iz2,jz0);
215 dx30 = _mm256_sub_pd(ix3,jx0);
216 dy30 = _mm256_sub_pd(iy3,jy0);
217 dz30 = _mm256_sub_pd(iz3,jz0);
219 /* Calculate squared distance and things based on it */
220 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
221 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
222 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
223 rsq30 = gmx_mm256_calc_rsq_pd(dx30,dy30,dz30);
225 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
226 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
227 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
228 rinv30 = gmx_mm256_invsqrt_pd(rsq30);
230 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
231 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
232 rinvsq30 = _mm256_mul_pd(rinv30,rinv30);
234 /* Load parameters for j particles */
235 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
236 charge+jnrC+0,charge+jnrD+0);
237 vdwjidx0A = 2*vdwtype[jnrA+0];
238 vdwjidx0B = 2*vdwtype[jnrB+0];
239 vdwjidx0C = 2*vdwtype[jnrC+0];
240 vdwjidx0D = 2*vdwtype[jnrD+0];
242 fjx0 = _mm256_setzero_pd();
243 fjy0 = _mm256_setzero_pd();
244 fjz0 = _mm256_setzero_pd();
246 /**************************
247 * CALCULATE INTERACTIONS *
248 **************************/
250 r00 = _mm256_mul_pd(rsq00,rinv00);
252 /* Compute parameters for interactions between i and j atoms */
253 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
254 vdwioffsetptr0+vdwjidx0B,
255 vdwioffsetptr0+vdwjidx0C,
256 vdwioffsetptr0+vdwjidx0D,
259 /* Calculate table index by multiplying r with table scale and truncate to integer */
260 rt = _mm256_mul_pd(r00,vftabscale);
261 vfitab = _mm256_cvttpd_epi32(rt);
262 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
263 vfitab = _mm_slli_epi32(vfitab,3);
265 /* CUBIC SPLINE TABLE DISPERSION */
266 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
267 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
268 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
269 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
270 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
271 Heps = _mm256_mul_pd(vfeps,H);
272 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
273 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
274 vvdw6 = _mm256_mul_pd(c6_00,VV);
275 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
276 fvdw6 = _mm256_mul_pd(c6_00,FF);
278 /* CUBIC SPLINE TABLE REPULSION */
279 vfitab = _mm_add_epi32(vfitab,ifour);
280 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
281 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
282 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
283 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
284 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
285 Heps = _mm256_mul_pd(vfeps,H);
286 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
287 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
288 vvdw12 = _mm256_mul_pd(c12_00,VV);
289 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
290 fvdw12 = _mm256_mul_pd(c12_00,FF);
291 vvdw = _mm256_add_pd(vvdw12,vvdw6);
292 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
294 /* Update potential sum for this i atom from the interaction with this j atom. */
295 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
299 /* Calculate temporary vectorial force */
300 tx = _mm256_mul_pd(fscal,dx00);
301 ty = _mm256_mul_pd(fscal,dy00);
302 tz = _mm256_mul_pd(fscal,dz00);
304 /* Update vectorial force */
305 fix0 = _mm256_add_pd(fix0,tx);
306 fiy0 = _mm256_add_pd(fiy0,ty);
307 fiz0 = _mm256_add_pd(fiz0,tz);
309 fjx0 = _mm256_add_pd(fjx0,tx);
310 fjy0 = _mm256_add_pd(fjy0,ty);
311 fjz0 = _mm256_add_pd(fjz0,tz);
313 /**************************
314 * CALCULATE INTERACTIONS *
315 **************************/
317 if (gmx_mm256_any_lt(rsq10,rcutoff2))
320 /* Compute parameters for interactions between i and j atoms */
321 qq10 = _mm256_mul_pd(iq1,jq0);
323 /* REACTION-FIELD ELECTROSTATICS */
324 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
325 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
327 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
329 /* Update potential sum for this i atom from the interaction with this j atom. */
330 velec = _mm256_and_pd(velec,cutoff_mask);
331 velecsum = _mm256_add_pd(velecsum,velec);
335 fscal = _mm256_and_pd(fscal,cutoff_mask);
337 /* Calculate temporary vectorial force */
338 tx = _mm256_mul_pd(fscal,dx10);
339 ty = _mm256_mul_pd(fscal,dy10);
340 tz = _mm256_mul_pd(fscal,dz10);
342 /* Update vectorial force */
343 fix1 = _mm256_add_pd(fix1,tx);
344 fiy1 = _mm256_add_pd(fiy1,ty);
345 fiz1 = _mm256_add_pd(fiz1,tz);
347 fjx0 = _mm256_add_pd(fjx0,tx);
348 fjy0 = _mm256_add_pd(fjy0,ty);
349 fjz0 = _mm256_add_pd(fjz0,tz);
353 /**************************
354 * CALCULATE INTERACTIONS *
355 **************************/
357 if (gmx_mm256_any_lt(rsq20,rcutoff2))
360 /* Compute parameters for interactions between i and j atoms */
361 qq20 = _mm256_mul_pd(iq2,jq0);
363 /* REACTION-FIELD ELECTROSTATICS */
364 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
365 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
367 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
369 /* Update potential sum for this i atom from the interaction with this j atom. */
370 velec = _mm256_and_pd(velec,cutoff_mask);
371 velecsum = _mm256_add_pd(velecsum,velec);
375 fscal = _mm256_and_pd(fscal,cutoff_mask);
377 /* Calculate temporary vectorial force */
378 tx = _mm256_mul_pd(fscal,dx20);
379 ty = _mm256_mul_pd(fscal,dy20);
380 tz = _mm256_mul_pd(fscal,dz20);
382 /* Update vectorial force */
383 fix2 = _mm256_add_pd(fix2,tx);
384 fiy2 = _mm256_add_pd(fiy2,ty);
385 fiz2 = _mm256_add_pd(fiz2,tz);
387 fjx0 = _mm256_add_pd(fjx0,tx);
388 fjy0 = _mm256_add_pd(fjy0,ty);
389 fjz0 = _mm256_add_pd(fjz0,tz);
393 /**************************
394 * CALCULATE INTERACTIONS *
395 **************************/
397 if (gmx_mm256_any_lt(rsq30,rcutoff2))
400 /* Compute parameters for interactions between i and j atoms */
401 qq30 = _mm256_mul_pd(iq3,jq0);
403 /* REACTION-FIELD ELECTROSTATICS */
404 velec = _mm256_mul_pd(qq30,_mm256_sub_pd(_mm256_add_pd(rinv30,_mm256_mul_pd(krf,rsq30)),crf));
405 felec = _mm256_mul_pd(qq30,_mm256_sub_pd(_mm256_mul_pd(rinv30,rinvsq30),krf2));
407 cutoff_mask = _mm256_cmp_pd(rsq30,rcutoff2,_CMP_LT_OQ);
409 /* Update potential sum for this i atom from the interaction with this j atom. */
410 velec = _mm256_and_pd(velec,cutoff_mask);
411 velecsum = _mm256_add_pd(velecsum,velec);
415 fscal = _mm256_and_pd(fscal,cutoff_mask);
417 /* Calculate temporary vectorial force */
418 tx = _mm256_mul_pd(fscal,dx30);
419 ty = _mm256_mul_pd(fscal,dy30);
420 tz = _mm256_mul_pd(fscal,dz30);
422 /* Update vectorial force */
423 fix3 = _mm256_add_pd(fix3,tx);
424 fiy3 = _mm256_add_pd(fiy3,ty);
425 fiz3 = _mm256_add_pd(fiz3,tz);
427 fjx0 = _mm256_add_pd(fjx0,tx);
428 fjy0 = _mm256_add_pd(fjy0,ty);
429 fjz0 = _mm256_add_pd(fjz0,tz);
433 fjptrA = f+j_coord_offsetA;
434 fjptrB = f+j_coord_offsetB;
435 fjptrC = f+j_coord_offsetC;
436 fjptrD = f+j_coord_offsetD;
438 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
440 /* Inner loop uses 167 flops */
446 /* Get j neighbor index, and coordinate index */
447 jnrlistA = jjnr[jidx];
448 jnrlistB = jjnr[jidx+1];
449 jnrlistC = jjnr[jidx+2];
450 jnrlistD = jjnr[jidx+3];
451 /* Sign of each element will be negative for non-real atoms.
452 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
453 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
455 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
457 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
458 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
459 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
461 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
462 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
463 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
464 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
465 j_coord_offsetA = DIM*jnrA;
466 j_coord_offsetB = DIM*jnrB;
467 j_coord_offsetC = DIM*jnrC;
468 j_coord_offsetD = DIM*jnrD;
470 /* load j atom coordinates */
471 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
472 x+j_coord_offsetC,x+j_coord_offsetD,
475 /* Calculate displacement vector */
476 dx00 = _mm256_sub_pd(ix0,jx0);
477 dy00 = _mm256_sub_pd(iy0,jy0);
478 dz00 = _mm256_sub_pd(iz0,jz0);
479 dx10 = _mm256_sub_pd(ix1,jx0);
480 dy10 = _mm256_sub_pd(iy1,jy0);
481 dz10 = _mm256_sub_pd(iz1,jz0);
482 dx20 = _mm256_sub_pd(ix2,jx0);
483 dy20 = _mm256_sub_pd(iy2,jy0);
484 dz20 = _mm256_sub_pd(iz2,jz0);
485 dx30 = _mm256_sub_pd(ix3,jx0);
486 dy30 = _mm256_sub_pd(iy3,jy0);
487 dz30 = _mm256_sub_pd(iz3,jz0);
489 /* Calculate squared distance and things based on it */
490 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
491 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
492 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
493 rsq30 = gmx_mm256_calc_rsq_pd(dx30,dy30,dz30);
495 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
496 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
497 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
498 rinv30 = gmx_mm256_invsqrt_pd(rsq30);
500 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
501 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
502 rinvsq30 = _mm256_mul_pd(rinv30,rinv30);
504 /* Load parameters for j particles */
505 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
506 charge+jnrC+0,charge+jnrD+0);
507 vdwjidx0A = 2*vdwtype[jnrA+0];
508 vdwjidx0B = 2*vdwtype[jnrB+0];
509 vdwjidx0C = 2*vdwtype[jnrC+0];
510 vdwjidx0D = 2*vdwtype[jnrD+0];
512 fjx0 = _mm256_setzero_pd();
513 fjy0 = _mm256_setzero_pd();
514 fjz0 = _mm256_setzero_pd();
516 /**************************
517 * CALCULATE INTERACTIONS *
518 **************************/
520 r00 = _mm256_mul_pd(rsq00,rinv00);
521 r00 = _mm256_andnot_pd(dummy_mask,r00);
523 /* Compute parameters for interactions between i and j atoms */
524 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
525 vdwioffsetptr0+vdwjidx0B,
526 vdwioffsetptr0+vdwjidx0C,
527 vdwioffsetptr0+vdwjidx0D,
530 /* Calculate table index by multiplying r with table scale and truncate to integer */
531 rt = _mm256_mul_pd(r00,vftabscale);
532 vfitab = _mm256_cvttpd_epi32(rt);
533 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
534 vfitab = _mm_slli_epi32(vfitab,3);
536 /* CUBIC SPLINE TABLE DISPERSION */
537 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
538 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
539 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
540 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
541 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
542 Heps = _mm256_mul_pd(vfeps,H);
543 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
544 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
545 vvdw6 = _mm256_mul_pd(c6_00,VV);
546 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
547 fvdw6 = _mm256_mul_pd(c6_00,FF);
549 /* CUBIC SPLINE TABLE REPULSION */
550 vfitab = _mm_add_epi32(vfitab,ifour);
551 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
552 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
553 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
554 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
555 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
556 Heps = _mm256_mul_pd(vfeps,H);
557 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
558 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
559 vvdw12 = _mm256_mul_pd(c12_00,VV);
560 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
561 fvdw12 = _mm256_mul_pd(c12_00,FF);
562 vvdw = _mm256_add_pd(vvdw12,vvdw6);
563 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
565 /* Update potential sum for this i atom from the interaction with this j atom. */
566 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
567 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
571 fscal = _mm256_andnot_pd(dummy_mask,fscal);
573 /* Calculate temporary vectorial force */
574 tx = _mm256_mul_pd(fscal,dx00);
575 ty = _mm256_mul_pd(fscal,dy00);
576 tz = _mm256_mul_pd(fscal,dz00);
578 /* Update vectorial force */
579 fix0 = _mm256_add_pd(fix0,tx);
580 fiy0 = _mm256_add_pd(fiy0,ty);
581 fiz0 = _mm256_add_pd(fiz0,tz);
583 fjx0 = _mm256_add_pd(fjx0,tx);
584 fjy0 = _mm256_add_pd(fjy0,ty);
585 fjz0 = _mm256_add_pd(fjz0,tz);
587 /**************************
588 * CALCULATE INTERACTIONS *
589 **************************/
591 if (gmx_mm256_any_lt(rsq10,rcutoff2))
594 /* Compute parameters for interactions between i and j atoms */
595 qq10 = _mm256_mul_pd(iq1,jq0);
597 /* REACTION-FIELD ELECTROSTATICS */
598 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
599 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
601 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
603 /* Update potential sum for this i atom from the interaction with this j atom. */
604 velec = _mm256_and_pd(velec,cutoff_mask);
605 velec = _mm256_andnot_pd(dummy_mask,velec);
606 velecsum = _mm256_add_pd(velecsum,velec);
610 fscal = _mm256_and_pd(fscal,cutoff_mask);
612 fscal = _mm256_andnot_pd(dummy_mask,fscal);
614 /* Calculate temporary vectorial force */
615 tx = _mm256_mul_pd(fscal,dx10);
616 ty = _mm256_mul_pd(fscal,dy10);
617 tz = _mm256_mul_pd(fscal,dz10);
619 /* Update vectorial force */
620 fix1 = _mm256_add_pd(fix1,tx);
621 fiy1 = _mm256_add_pd(fiy1,ty);
622 fiz1 = _mm256_add_pd(fiz1,tz);
624 fjx0 = _mm256_add_pd(fjx0,tx);
625 fjy0 = _mm256_add_pd(fjy0,ty);
626 fjz0 = _mm256_add_pd(fjz0,tz);
630 /**************************
631 * CALCULATE INTERACTIONS *
632 **************************/
634 if (gmx_mm256_any_lt(rsq20,rcutoff2))
637 /* Compute parameters for interactions between i and j atoms */
638 qq20 = _mm256_mul_pd(iq2,jq0);
640 /* REACTION-FIELD ELECTROSTATICS */
641 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
642 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
644 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
646 /* Update potential sum for this i atom from the interaction with this j atom. */
647 velec = _mm256_and_pd(velec,cutoff_mask);
648 velec = _mm256_andnot_pd(dummy_mask,velec);
649 velecsum = _mm256_add_pd(velecsum,velec);
653 fscal = _mm256_and_pd(fscal,cutoff_mask);
655 fscal = _mm256_andnot_pd(dummy_mask,fscal);
657 /* Calculate temporary vectorial force */
658 tx = _mm256_mul_pd(fscal,dx20);
659 ty = _mm256_mul_pd(fscal,dy20);
660 tz = _mm256_mul_pd(fscal,dz20);
662 /* Update vectorial force */
663 fix2 = _mm256_add_pd(fix2,tx);
664 fiy2 = _mm256_add_pd(fiy2,ty);
665 fiz2 = _mm256_add_pd(fiz2,tz);
667 fjx0 = _mm256_add_pd(fjx0,tx);
668 fjy0 = _mm256_add_pd(fjy0,ty);
669 fjz0 = _mm256_add_pd(fjz0,tz);
673 /**************************
674 * CALCULATE INTERACTIONS *
675 **************************/
677 if (gmx_mm256_any_lt(rsq30,rcutoff2))
680 /* Compute parameters for interactions between i and j atoms */
681 qq30 = _mm256_mul_pd(iq3,jq0);
683 /* REACTION-FIELD ELECTROSTATICS */
684 velec = _mm256_mul_pd(qq30,_mm256_sub_pd(_mm256_add_pd(rinv30,_mm256_mul_pd(krf,rsq30)),crf));
685 felec = _mm256_mul_pd(qq30,_mm256_sub_pd(_mm256_mul_pd(rinv30,rinvsq30),krf2));
687 cutoff_mask = _mm256_cmp_pd(rsq30,rcutoff2,_CMP_LT_OQ);
689 /* Update potential sum for this i atom from the interaction with this j atom. */
690 velec = _mm256_and_pd(velec,cutoff_mask);
691 velec = _mm256_andnot_pd(dummy_mask,velec);
692 velecsum = _mm256_add_pd(velecsum,velec);
696 fscal = _mm256_and_pd(fscal,cutoff_mask);
698 fscal = _mm256_andnot_pd(dummy_mask,fscal);
700 /* Calculate temporary vectorial force */
701 tx = _mm256_mul_pd(fscal,dx30);
702 ty = _mm256_mul_pd(fscal,dy30);
703 tz = _mm256_mul_pd(fscal,dz30);
705 /* Update vectorial force */
706 fix3 = _mm256_add_pd(fix3,tx);
707 fiy3 = _mm256_add_pd(fiy3,ty);
708 fiz3 = _mm256_add_pd(fiz3,tz);
710 fjx0 = _mm256_add_pd(fjx0,tx);
711 fjy0 = _mm256_add_pd(fjy0,ty);
712 fjz0 = _mm256_add_pd(fjz0,tz);
716 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
717 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
718 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
719 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
721 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
723 /* Inner loop uses 168 flops */
726 /* End of innermost loop */
728 gmx_mm256_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
729 f+i_coord_offset,fshift+i_shift_offset);
732 /* Update potential energies */
733 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
734 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
736 /* Increment number of inner iterations */
737 inneriter += j_index_end - j_index_start;
739 /* Outer loop uses 26 flops */
742 /* Increment number of outer iterations */
745 /* Update outer/inner flops */
747 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*168);
750 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_F_avx_256_double
751 * Electrostatics interaction: ReactionField
752 * VdW interaction: CubicSplineTable
753 * Geometry: Water4-Particle
754 * Calculate force/pot: Force
757 nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_F_avx_256_double
758 (t_nblist * gmx_restrict nlist,
759 rvec * gmx_restrict xx,
760 rvec * gmx_restrict ff,
761 t_forcerec * gmx_restrict fr,
762 t_mdatoms * gmx_restrict mdatoms,
763 nb_kernel_data_t * gmx_restrict kernel_data,
764 t_nrnb * gmx_restrict nrnb)
766 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
767 * just 0 for non-waters.
768 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
769 * jnr indices corresponding to data put in the four positions in the SIMD register.
771 int i_shift_offset,i_coord_offset,outeriter,inneriter;
772 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
773 int jnrA,jnrB,jnrC,jnrD;
774 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
775 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
776 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
777 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
779 real *shiftvec,*fshift,*x,*f;
780 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
782 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
783 real * vdwioffsetptr0;
784 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
785 real * vdwioffsetptr1;
786 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
787 real * vdwioffsetptr2;
788 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
789 real * vdwioffsetptr3;
790 __m256d ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
791 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
792 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
793 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
794 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
795 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
796 __m256d dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
797 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
800 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
803 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
804 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
806 __m128i ifour = _mm_set1_epi32(4);
807 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
809 __m256d dummy_mask,cutoff_mask;
810 __m128 tmpmask0,tmpmask1;
811 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
812 __m256d one = _mm256_set1_pd(1.0);
813 __m256d two = _mm256_set1_pd(2.0);
819 jindex = nlist->jindex;
821 shiftidx = nlist->shift;
823 shiftvec = fr->shift_vec[0];
824 fshift = fr->fshift[0];
825 facel = _mm256_set1_pd(fr->epsfac);
826 charge = mdatoms->chargeA;
827 krf = _mm256_set1_pd(fr->ic->k_rf);
828 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
829 crf = _mm256_set1_pd(fr->ic->c_rf);
830 nvdwtype = fr->ntype;
832 vdwtype = mdatoms->typeA;
834 vftab = kernel_data->table_vdw->data;
835 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
837 /* Setup water-specific parameters */
838 inr = nlist->iinr[0];
839 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
840 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
841 iq3 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+3]));
842 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
844 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
845 rcutoff_scalar = fr->rcoulomb;
846 rcutoff = _mm256_set1_pd(rcutoff_scalar);
847 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
849 /* Avoid stupid compiler warnings */
850 jnrA = jnrB = jnrC = jnrD = 0;
859 for(iidx=0;iidx<4*DIM;iidx++)
864 /* Start outer loop over neighborlists */
865 for(iidx=0; iidx<nri; iidx++)
867 /* Load shift vector for this list */
868 i_shift_offset = DIM*shiftidx[iidx];
870 /* Load limits for loop over neighbors */
871 j_index_start = jindex[iidx];
872 j_index_end = jindex[iidx+1];
874 /* Get outer coordinate index */
876 i_coord_offset = DIM*inr;
878 /* Load i particle coords and add shift vector */
879 gmx_mm256_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
880 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
882 fix0 = _mm256_setzero_pd();
883 fiy0 = _mm256_setzero_pd();
884 fiz0 = _mm256_setzero_pd();
885 fix1 = _mm256_setzero_pd();
886 fiy1 = _mm256_setzero_pd();
887 fiz1 = _mm256_setzero_pd();
888 fix2 = _mm256_setzero_pd();
889 fiy2 = _mm256_setzero_pd();
890 fiz2 = _mm256_setzero_pd();
891 fix3 = _mm256_setzero_pd();
892 fiy3 = _mm256_setzero_pd();
893 fiz3 = _mm256_setzero_pd();
895 /* Start inner kernel loop */
896 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
899 /* Get j neighbor index, and coordinate index */
904 j_coord_offsetA = DIM*jnrA;
905 j_coord_offsetB = DIM*jnrB;
906 j_coord_offsetC = DIM*jnrC;
907 j_coord_offsetD = DIM*jnrD;
909 /* load j atom coordinates */
910 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
911 x+j_coord_offsetC,x+j_coord_offsetD,
914 /* Calculate displacement vector */
915 dx00 = _mm256_sub_pd(ix0,jx0);
916 dy00 = _mm256_sub_pd(iy0,jy0);
917 dz00 = _mm256_sub_pd(iz0,jz0);
918 dx10 = _mm256_sub_pd(ix1,jx0);
919 dy10 = _mm256_sub_pd(iy1,jy0);
920 dz10 = _mm256_sub_pd(iz1,jz0);
921 dx20 = _mm256_sub_pd(ix2,jx0);
922 dy20 = _mm256_sub_pd(iy2,jy0);
923 dz20 = _mm256_sub_pd(iz2,jz0);
924 dx30 = _mm256_sub_pd(ix3,jx0);
925 dy30 = _mm256_sub_pd(iy3,jy0);
926 dz30 = _mm256_sub_pd(iz3,jz0);
928 /* Calculate squared distance and things based on it */
929 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
930 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
931 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
932 rsq30 = gmx_mm256_calc_rsq_pd(dx30,dy30,dz30);
934 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
935 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
936 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
937 rinv30 = gmx_mm256_invsqrt_pd(rsq30);
939 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
940 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
941 rinvsq30 = _mm256_mul_pd(rinv30,rinv30);
943 /* Load parameters for j particles */
944 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
945 charge+jnrC+0,charge+jnrD+0);
946 vdwjidx0A = 2*vdwtype[jnrA+0];
947 vdwjidx0B = 2*vdwtype[jnrB+0];
948 vdwjidx0C = 2*vdwtype[jnrC+0];
949 vdwjidx0D = 2*vdwtype[jnrD+0];
951 fjx0 = _mm256_setzero_pd();
952 fjy0 = _mm256_setzero_pd();
953 fjz0 = _mm256_setzero_pd();
955 /**************************
956 * CALCULATE INTERACTIONS *
957 **************************/
959 r00 = _mm256_mul_pd(rsq00,rinv00);
961 /* Compute parameters for interactions between i and j atoms */
962 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
963 vdwioffsetptr0+vdwjidx0B,
964 vdwioffsetptr0+vdwjidx0C,
965 vdwioffsetptr0+vdwjidx0D,
968 /* Calculate table index by multiplying r with table scale and truncate to integer */
969 rt = _mm256_mul_pd(r00,vftabscale);
970 vfitab = _mm256_cvttpd_epi32(rt);
971 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
972 vfitab = _mm_slli_epi32(vfitab,3);
974 /* CUBIC SPLINE TABLE DISPERSION */
975 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
976 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
977 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
978 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
979 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
980 Heps = _mm256_mul_pd(vfeps,H);
981 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
982 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
983 fvdw6 = _mm256_mul_pd(c6_00,FF);
985 /* CUBIC SPLINE TABLE REPULSION */
986 vfitab = _mm_add_epi32(vfitab,ifour);
987 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
988 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
989 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
990 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
991 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
992 Heps = _mm256_mul_pd(vfeps,H);
993 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
994 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
995 fvdw12 = _mm256_mul_pd(c12_00,FF);
996 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
1000 /* Calculate temporary vectorial force */
1001 tx = _mm256_mul_pd(fscal,dx00);
1002 ty = _mm256_mul_pd(fscal,dy00);
1003 tz = _mm256_mul_pd(fscal,dz00);
1005 /* Update vectorial force */
1006 fix0 = _mm256_add_pd(fix0,tx);
1007 fiy0 = _mm256_add_pd(fiy0,ty);
1008 fiz0 = _mm256_add_pd(fiz0,tz);
1010 fjx0 = _mm256_add_pd(fjx0,tx);
1011 fjy0 = _mm256_add_pd(fjy0,ty);
1012 fjz0 = _mm256_add_pd(fjz0,tz);
1014 /**************************
1015 * CALCULATE INTERACTIONS *
1016 **************************/
1018 if (gmx_mm256_any_lt(rsq10,rcutoff2))
1021 /* Compute parameters for interactions between i and j atoms */
1022 qq10 = _mm256_mul_pd(iq1,jq0);
1024 /* REACTION-FIELD ELECTROSTATICS */
1025 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
1027 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
1031 fscal = _mm256_and_pd(fscal,cutoff_mask);
1033 /* Calculate temporary vectorial force */
1034 tx = _mm256_mul_pd(fscal,dx10);
1035 ty = _mm256_mul_pd(fscal,dy10);
1036 tz = _mm256_mul_pd(fscal,dz10);
1038 /* Update vectorial force */
1039 fix1 = _mm256_add_pd(fix1,tx);
1040 fiy1 = _mm256_add_pd(fiy1,ty);
1041 fiz1 = _mm256_add_pd(fiz1,tz);
1043 fjx0 = _mm256_add_pd(fjx0,tx);
1044 fjy0 = _mm256_add_pd(fjy0,ty);
1045 fjz0 = _mm256_add_pd(fjz0,tz);
1049 /**************************
1050 * CALCULATE INTERACTIONS *
1051 **************************/
1053 if (gmx_mm256_any_lt(rsq20,rcutoff2))
1056 /* Compute parameters for interactions between i and j atoms */
1057 qq20 = _mm256_mul_pd(iq2,jq0);
1059 /* REACTION-FIELD ELECTROSTATICS */
1060 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
1062 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1066 fscal = _mm256_and_pd(fscal,cutoff_mask);
1068 /* Calculate temporary vectorial force */
1069 tx = _mm256_mul_pd(fscal,dx20);
1070 ty = _mm256_mul_pd(fscal,dy20);
1071 tz = _mm256_mul_pd(fscal,dz20);
1073 /* Update vectorial force */
1074 fix2 = _mm256_add_pd(fix2,tx);
1075 fiy2 = _mm256_add_pd(fiy2,ty);
1076 fiz2 = _mm256_add_pd(fiz2,tz);
1078 fjx0 = _mm256_add_pd(fjx0,tx);
1079 fjy0 = _mm256_add_pd(fjy0,ty);
1080 fjz0 = _mm256_add_pd(fjz0,tz);
1084 /**************************
1085 * CALCULATE INTERACTIONS *
1086 **************************/
1088 if (gmx_mm256_any_lt(rsq30,rcutoff2))
1091 /* Compute parameters for interactions between i and j atoms */
1092 qq30 = _mm256_mul_pd(iq3,jq0);
1094 /* REACTION-FIELD ELECTROSTATICS */
1095 felec = _mm256_mul_pd(qq30,_mm256_sub_pd(_mm256_mul_pd(rinv30,rinvsq30),krf2));
1097 cutoff_mask = _mm256_cmp_pd(rsq30,rcutoff2,_CMP_LT_OQ);
1101 fscal = _mm256_and_pd(fscal,cutoff_mask);
1103 /* Calculate temporary vectorial force */
1104 tx = _mm256_mul_pd(fscal,dx30);
1105 ty = _mm256_mul_pd(fscal,dy30);
1106 tz = _mm256_mul_pd(fscal,dz30);
1108 /* Update vectorial force */
1109 fix3 = _mm256_add_pd(fix3,tx);
1110 fiy3 = _mm256_add_pd(fiy3,ty);
1111 fiz3 = _mm256_add_pd(fiz3,tz);
1113 fjx0 = _mm256_add_pd(fjx0,tx);
1114 fjy0 = _mm256_add_pd(fjy0,ty);
1115 fjz0 = _mm256_add_pd(fjz0,tz);
1119 fjptrA = f+j_coord_offsetA;
1120 fjptrB = f+j_coord_offsetB;
1121 fjptrC = f+j_coord_offsetC;
1122 fjptrD = f+j_coord_offsetD;
1124 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1126 /* Inner loop uses 141 flops */
1129 if(jidx<j_index_end)
1132 /* Get j neighbor index, and coordinate index */
1133 jnrlistA = jjnr[jidx];
1134 jnrlistB = jjnr[jidx+1];
1135 jnrlistC = jjnr[jidx+2];
1136 jnrlistD = jjnr[jidx+3];
1137 /* Sign of each element will be negative for non-real atoms.
1138 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1139 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
1141 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1143 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
1144 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
1145 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
1147 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1148 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1149 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1150 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1151 j_coord_offsetA = DIM*jnrA;
1152 j_coord_offsetB = DIM*jnrB;
1153 j_coord_offsetC = DIM*jnrC;
1154 j_coord_offsetD = DIM*jnrD;
1156 /* load j atom coordinates */
1157 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1158 x+j_coord_offsetC,x+j_coord_offsetD,
1161 /* Calculate displacement vector */
1162 dx00 = _mm256_sub_pd(ix0,jx0);
1163 dy00 = _mm256_sub_pd(iy0,jy0);
1164 dz00 = _mm256_sub_pd(iz0,jz0);
1165 dx10 = _mm256_sub_pd(ix1,jx0);
1166 dy10 = _mm256_sub_pd(iy1,jy0);
1167 dz10 = _mm256_sub_pd(iz1,jz0);
1168 dx20 = _mm256_sub_pd(ix2,jx0);
1169 dy20 = _mm256_sub_pd(iy2,jy0);
1170 dz20 = _mm256_sub_pd(iz2,jz0);
1171 dx30 = _mm256_sub_pd(ix3,jx0);
1172 dy30 = _mm256_sub_pd(iy3,jy0);
1173 dz30 = _mm256_sub_pd(iz3,jz0);
1175 /* Calculate squared distance and things based on it */
1176 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
1177 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
1178 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
1179 rsq30 = gmx_mm256_calc_rsq_pd(dx30,dy30,dz30);
1181 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
1182 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
1183 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
1184 rinv30 = gmx_mm256_invsqrt_pd(rsq30);
1186 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
1187 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
1188 rinvsq30 = _mm256_mul_pd(rinv30,rinv30);
1190 /* Load parameters for j particles */
1191 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
1192 charge+jnrC+0,charge+jnrD+0);
1193 vdwjidx0A = 2*vdwtype[jnrA+0];
1194 vdwjidx0B = 2*vdwtype[jnrB+0];
1195 vdwjidx0C = 2*vdwtype[jnrC+0];
1196 vdwjidx0D = 2*vdwtype[jnrD+0];
1198 fjx0 = _mm256_setzero_pd();
1199 fjy0 = _mm256_setzero_pd();
1200 fjz0 = _mm256_setzero_pd();
1202 /**************************
1203 * CALCULATE INTERACTIONS *
1204 **************************/
1206 r00 = _mm256_mul_pd(rsq00,rinv00);
1207 r00 = _mm256_andnot_pd(dummy_mask,r00);
1209 /* Compute parameters for interactions between i and j atoms */
1210 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
1211 vdwioffsetptr0+vdwjidx0B,
1212 vdwioffsetptr0+vdwjidx0C,
1213 vdwioffsetptr0+vdwjidx0D,
1216 /* Calculate table index by multiplying r with table scale and truncate to integer */
1217 rt = _mm256_mul_pd(r00,vftabscale);
1218 vfitab = _mm256_cvttpd_epi32(rt);
1219 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
1220 vfitab = _mm_slli_epi32(vfitab,3);
1222 /* CUBIC SPLINE TABLE DISPERSION */
1223 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1224 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1225 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
1226 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
1227 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
1228 Heps = _mm256_mul_pd(vfeps,H);
1229 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
1230 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
1231 fvdw6 = _mm256_mul_pd(c6_00,FF);
1233 /* CUBIC SPLINE TABLE REPULSION */
1234 vfitab = _mm_add_epi32(vfitab,ifour);
1235 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1236 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1237 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
1238 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
1239 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
1240 Heps = _mm256_mul_pd(vfeps,H);
1241 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
1242 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
1243 fvdw12 = _mm256_mul_pd(c12_00,FF);
1244 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
1248 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1250 /* Calculate temporary vectorial force */
1251 tx = _mm256_mul_pd(fscal,dx00);
1252 ty = _mm256_mul_pd(fscal,dy00);
1253 tz = _mm256_mul_pd(fscal,dz00);
1255 /* Update vectorial force */
1256 fix0 = _mm256_add_pd(fix0,tx);
1257 fiy0 = _mm256_add_pd(fiy0,ty);
1258 fiz0 = _mm256_add_pd(fiz0,tz);
1260 fjx0 = _mm256_add_pd(fjx0,tx);
1261 fjy0 = _mm256_add_pd(fjy0,ty);
1262 fjz0 = _mm256_add_pd(fjz0,tz);
1264 /**************************
1265 * CALCULATE INTERACTIONS *
1266 **************************/
1268 if (gmx_mm256_any_lt(rsq10,rcutoff2))
1271 /* Compute parameters for interactions between i and j atoms */
1272 qq10 = _mm256_mul_pd(iq1,jq0);
1274 /* REACTION-FIELD ELECTROSTATICS */
1275 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
1277 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
1281 fscal = _mm256_and_pd(fscal,cutoff_mask);
1283 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1285 /* Calculate temporary vectorial force */
1286 tx = _mm256_mul_pd(fscal,dx10);
1287 ty = _mm256_mul_pd(fscal,dy10);
1288 tz = _mm256_mul_pd(fscal,dz10);
1290 /* Update vectorial force */
1291 fix1 = _mm256_add_pd(fix1,tx);
1292 fiy1 = _mm256_add_pd(fiy1,ty);
1293 fiz1 = _mm256_add_pd(fiz1,tz);
1295 fjx0 = _mm256_add_pd(fjx0,tx);
1296 fjy0 = _mm256_add_pd(fjy0,ty);
1297 fjz0 = _mm256_add_pd(fjz0,tz);
1301 /**************************
1302 * CALCULATE INTERACTIONS *
1303 **************************/
1305 if (gmx_mm256_any_lt(rsq20,rcutoff2))
1308 /* Compute parameters for interactions between i and j atoms */
1309 qq20 = _mm256_mul_pd(iq2,jq0);
1311 /* REACTION-FIELD ELECTROSTATICS */
1312 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
1314 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1318 fscal = _mm256_and_pd(fscal,cutoff_mask);
1320 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1322 /* Calculate temporary vectorial force */
1323 tx = _mm256_mul_pd(fscal,dx20);
1324 ty = _mm256_mul_pd(fscal,dy20);
1325 tz = _mm256_mul_pd(fscal,dz20);
1327 /* Update vectorial force */
1328 fix2 = _mm256_add_pd(fix2,tx);
1329 fiy2 = _mm256_add_pd(fiy2,ty);
1330 fiz2 = _mm256_add_pd(fiz2,tz);
1332 fjx0 = _mm256_add_pd(fjx0,tx);
1333 fjy0 = _mm256_add_pd(fjy0,ty);
1334 fjz0 = _mm256_add_pd(fjz0,tz);
1338 /**************************
1339 * CALCULATE INTERACTIONS *
1340 **************************/
1342 if (gmx_mm256_any_lt(rsq30,rcutoff2))
1345 /* Compute parameters for interactions between i and j atoms */
1346 qq30 = _mm256_mul_pd(iq3,jq0);
1348 /* REACTION-FIELD ELECTROSTATICS */
1349 felec = _mm256_mul_pd(qq30,_mm256_sub_pd(_mm256_mul_pd(rinv30,rinvsq30),krf2));
1351 cutoff_mask = _mm256_cmp_pd(rsq30,rcutoff2,_CMP_LT_OQ);
1355 fscal = _mm256_and_pd(fscal,cutoff_mask);
1357 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1359 /* Calculate temporary vectorial force */
1360 tx = _mm256_mul_pd(fscal,dx30);
1361 ty = _mm256_mul_pd(fscal,dy30);
1362 tz = _mm256_mul_pd(fscal,dz30);
1364 /* Update vectorial force */
1365 fix3 = _mm256_add_pd(fix3,tx);
1366 fiy3 = _mm256_add_pd(fiy3,ty);
1367 fiz3 = _mm256_add_pd(fiz3,tz);
1369 fjx0 = _mm256_add_pd(fjx0,tx);
1370 fjy0 = _mm256_add_pd(fjy0,ty);
1371 fjz0 = _mm256_add_pd(fjz0,tz);
1375 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1376 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1377 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1378 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1380 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1382 /* Inner loop uses 142 flops */
1385 /* End of innermost loop */
1387 gmx_mm256_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1388 f+i_coord_offset,fshift+i_shift_offset);
1390 /* Increment number of inner iterations */
1391 inneriter += j_index_end - j_index_start;
1393 /* Outer loop uses 24 flops */
1396 /* Increment number of outer iterations */
1399 /* Update outer/inner flops */
1401 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*142);